Kai Ni's research while affiliated with Notre Dame College and other places

... The HDC operators -binding, bundling, and permutation -construct sets, associations, and sequences respectively, facilitating the interpretable creation and manipulation of complex objects for data representation, learning, and processing. For learning, an HDC model makes decisions by evaluating the similarity between query and model hypervectors [11,26,12,18,16]; for cognitive processing, an HDC model retrieves information directly over the hyperspace with HDC operators; it then decodes the information with similarity functions and the atomic hypervectors [9,22,30]. Recent work has shown great advantages of HDC in enhancing the cognitive capability of neural networks in an explainable fashion [7]: a neural network learns to encode and perform HDC-like composition of the data over Raven's Progressive Matrix, a visual reasoning task over the symbolic attributes of the objects, and significantly outperforms state-of-the-art pure DNN and neuro-symbolic AI solutions in both accuracy and efficiency scaling. While HDC serves as a natural bridge between neural models and symbolic reasoning, it faces some scalability issues when it comes to decoding, a crucial process for information retrieval. ...

... COPs are a class of problems where the decision maker is required to make a series of choices from a limited set of options, and the goal is to find the optimal set of combinations that achieves the best outcomes. Combinatorial optimization plays a central role in tackling some of the most challenging problems across diverse domains [2,3,4,5,6], from logistics and transportation to resource allocation and scheduling. For example, logistics companies use combinatorial optimization to optimize delivery routes and schedules. ...

... However, there are applications that need to apply the V READ constantly, which brings serious concern over the state stability. A noteworthy example is FeFET based reconfigurable computing, where the FeFET is exploited as a compact nonvolatile 3 switch as shown in Fig.1a 13,14 . By combining the configuration memory and a switch together, a FeFET based nonvolatile switch can implement the routing network in a field programmable gate array (FPGA) by constructing the connection box (CB) and switch box (SB). ...

... Although JJ and SQUID have been fundamental components of superconducting circuits and systems [7,11], circuits based on these devices encounter several challenges, including limited cascadability, fabrication complexities, and poor scalability [2,18]. Three-terminal cryotron devices, demonstrating channel switching between superconducting and resistive states controlled by gate current, were proposed in response to these challenges [18]. ...